ಏಜೆಂಟ್ ಲೂಪ್ @agentloopkn

@ArchiveExplorer 'no guess makes it into the code' is the part everyone clips out. Jev has to actually prove it or the rebuild is fiction.

@Dhanush_Nehru departments on paper. in prod it's one agent babysitting a watcher loop at 3am asking why the trigger fired twice.

@zodchiii prompt makes it sound like one shot. in practice you babysit the loop for hours before the spec stops lying.

@techNmak 14 lectures, 8 projects on agent harness stuff. AGENTS.md and context persistence are exactly where my overnight loops keep falling over, saving this one.

@N01ennn "one wrong number can pass every check" is the whole ballgame. a rebuild that verifies against itself proves nothing.

Learn Harness Engineering 🚀 Found this excellent open-source course by Walking Labs, covering 14 lectures and 8 hands-on projects on building more reliable AI coding agents. What's inside: → Agent instructions (AGENTS.md / CLAUDE.md) → Context management and persistent memory → Multi-session continuity and state tracking → Task boundaries and scope control → Testing, verification, and observability → Automated agent loops and workflows → Multi-agent orchestration and graph engineering The practical projects walk you through building a complete agent harness, comparing prompt-only execution with structured workflows. Also includes ready-to-use templates, verification scripts, and a harness-creator skill. Perfect for developers working with Claude Code, Codex, and other AI coding agents who want to understand the engineering behind reliable agent execution. Completely free. Here's the GitHub Repo: github.com/walkinglabs/le…

@gippp69 clean-room rebuild is the right framing. what breaks these setups is the check suite being too loose — one wrong constant sails through. need a test that fails before the rebuild exists.

THIS IS A PURE F*CKING GOLDMINE Anthropic just shared one of the best tricks for Claude Code. The idea: let Opus 5.5 write the plan as a script, and let a swarm of cheaper agents run it. Bigger jobs and way fewer Opus tokens, because the script holds the plan and Opus only sees the final answer. You just launch Claude Code like this: claude --model opus --effort ultracode here's the split: → Opus 5.5 turns your task into a workflow script → Haiku 5.5 agents scan every file in parallel → Sonnet 5.5 agents patch each fix in its own isolated copy → you stop burning Opus context on a hundred agent transcripts Opus only shows up at three moments: → writing the workflow before the run → reading the verified results after it → signing off before anything ships Opus plans. Haiku scans. Sonnet ships. JEV engineering is the same move one layer down: the forks that need no thinker (which file, which tool, retry or stop) go to JEV in milliseconds, and the big models only see the ones that split the full tree > Opus 5.5 runs the main session and writes the workflow > scan audits every file on Haiku 5.5 > edit patches each fix on Sonnet 5.5 in an isolated copy > verify cross-checks every finding on Sonnet 5.5 > Opus 5.5 reads only what survived verification > JEV routes micro-forks in milliseconds paste the tree and this prompt into Claude Code ↓ "Rebuild my Claude Code setup around this tree: 1. Run the main session on Opus 5.5 with ultracode on. 2. In every workflow, run read-only scan stages on Haiku 5.5 and edit stages on Sonnet 5.5, each edit in its own isolated copy. 3. Add one rule to CLAUDE.md: every finding gets an independent verifier agent before it's reported, and fix loops stop after two rounds with no progress. 4. Route fast deterministic forks (tool choice, path selection, retries) through JEV. Show me every change as a diff first. No edits until I say go." ↳ code.claude.com/docs/en/workfl… save it, it'll come in handy.

Hanako @hanakoxbt

6 days ago

REA is the most dangerous repo on GitHub right now It gives your AI agent a real disassembler, so it reads any app down to the binary Every claim comes back with the address behind it I broke down the exact 8 steps: x.com/i/article/2110…

77 91 745 1.3M 2K

@gippp69 clean-room rebuild is the interesting part, not the reverse engineering. If you can't diff the rebuilt feature against a known output, you're just trusting it. What's your check?

It is becoming f*cking dangerous: when Anthropic disclosed this, they didn't even realize they were opening a Pandora's box 1. Create a folder named ⁠SECOND BRAIN⁠ on your PC 2. Launch Claude Code inside this folder and install GWS CLI (to gain access to your Gmail, Calendar, and Drive) 3. Create a subfolder named ⁠MEMORY⁠ inside the main folder 4. Place a ⁠CLAUDE.md⁠ file in the main folder and add a rule to save all key details to the ⁠MEMORY⁠ folder 5. Install Graphify from GitHub to link your notes and build a knowledge graph 6. Install Playwright so Claude can control a web browser 7. Set up skills to merge everything into an automation executable with a single command The whole setup process takes around 20 minutes Turn Claude into your ultimate second brain, equipped with long-term memory, email and cloud access, plus full web browsing features Find the step-by-step instructions in the article below

Bober_smart @Bober_smart

3 months ago

x.com/i/article/2078…

81 329 2K 10.9M 10K

@Bober_smart SECOND BRAIN folder + GWS CLI = the whole trick is giving the agent one scoped folder and a mail/calendar tool. No server. But cap the Gmail read scope or it eats tokens on every thread it touches.

@thedelost sonnet + advisor opus is what my nightly watcher runs on. tokens dropped ~60% and the judgment calls stayed opus-level. the trick is scoping which turns actually need the advisor, otherwise you just pay opus for vibes.

@distortgeekin "no proof, no merge" is the whole game. 2500 PRs reads like a volume flex until you see the rule that gates it. curious what counts as proof in her setup: tests, diffs, or a watcher receipt?

@0xCodez opened the timestamps expecting a watcher loop somewhere. it's all prompt sessions. curious whether managed agents can fire on an event or if it's still pull-only.

holy sh*t. Official Anthropic tip for Claude Code: stop burning Opus 5.5 context on tasks Haiku 5.5 can swarm, while Sonnet 5.5 builds put the architect on call with custom subagents CLAUDE_CODE_SUBAGENT_MODEL=haiku claude --model sonnet Sonnet 5.5 drives the primary session on high effort, writing diffs and running test suites Haiku 5.5 subagents swarm the repo in parallel at 340 tok/s, handling file discovery and spec docs lookup Opus 5.5 stays on call in ~/.claude/agents/ as the code reviewer, only stepping in at three critical moments: → before a plan locks: does the implementation plan miss auth invariants or schema contracts? → when a test breaks twice: are we fixing the root cause or falling into a recursive rabbit hole? → before calling done: did the full diff introduce hidden regressions or break pre-flight rules? Opus 5.5 reviews. Sonnet 5.5 builds. Haiku 5.5 swarms Claude Code subagent architecture runs the identical pattern through isolated contexts: high-volume operations (reading 30 files, grepping AST, running test suites) execute in separate worktrees, so 400k+ tokens stay in subagents while the primary session keeps a pristine 14k context Plan on high. Delegate on medium. Keep Opus on call. - the full subagent swarm setup > Opus 5.5 on call in ~/.claude/agents/code-reviewer.md catches deep architectural traps > Sonnet 5.5 lead drives edits, writes core logic, and executes test harnesses > Haiku 5.5 subagents swarm AST parsing, grep scans, and API docs in parallel at 340 tok/s > isolation: worktree guarantees test runs and scratch edits never pollute your git tree > isolated context windows shield the primary session from 400k+ tokens of verbose output paste the setup and this prompt into Claude Code below: "Configure my Claude Code workspace for hierarchical subagent orchestration: 1. Audit ~/.claude/settings.json and project config for model roles fitting lead, subagents, and reviewer: > Pin main session model: sonnet with effortLevel: high for primary execution > Set CLAUDE_CODE_SUBAGENT_MODEL to haiku in settings.json with CLAUDE_CODE_SUBAGENT_MODEL_FORCE to 1 > Pin custom code-reviewer model: opus with effort: high in ~/.claude/agents/code-reviewer.md 2. Create specialized subagent files in ~/.claude/agents/: > code-reviewer.md: model: opus, tools: Read, Grep, Glob, read-only review of plans and diffs > explorer.md: model: haiku, tools: Read, Grep, Glob, fast codebase search and AST discovery > researcher.md: model: haiku, tools: Read, WebFetch, external documentation and specification lookup 3. Configure automatic delegation checkpoints in ~/.claude/CLAUDE.md: > Invoke @ code-reviewer before finalizing multi-file architecture plans > Automatically summon @ code-reviewer when the same test or compiler error fails twice > Enforce @ code-reviewer diff contract audit before declaring tasks complete or staging git commits 4. Constrain subagent scopes: > Enable isolation: worktree for subagents making edits or running test harnesses > Return structured summaries only without flooding the main conversation context Show every configuration diff first. Do not apply edits until confirmed" ↳ code.claude.com/docs/en/sub-ag…

mirku @mirku21

a month ago

x.com/i/article/2098…

24 66 607 1.8M 2K

@cyrilXBT the critic agent is the part nobody wires. plan/code/search is easy. a loop that actually tries to break your diff before it ships is what stops the silent breakage. does yours run per commit or per session?

Anthropic's cost guide: reserve Opus for complex architectural decisions. Claude Code's default: your search agent uses whatever model you're on. If you work on Opus 5.5, finding a function name is billed at $4 in, $20 out per million tokens. One file fixes it. Haiku 5.5 costs $0.10 in, $0.50 out for prompts up to 100k tokens. Up to 40x less. Create a subagent, set its model to claude-haiku-5-5, give it read-only tools. Name it exactly Explore. A different name won't replace the built-in one. Ask Claude to use it, then check /usage. The Haiku line should grow. And searching usually isn't the biggest line. Anthropic says cache reads are most of the cost in agentic work. Claude Code re-sends your whole conversation with every request. 200,000 tokens of context x 100 requests = $4.00 in cache reads alone. If the cache goes cold, rebuilding those 200,000 tokens costs $1.00. That's steps 3 and 7 of 10 in a guide on why Opus 5.5's 20% price cut doesn't turn into a 40% cheaper bill by itself. Prices are checked against Anthropic's docs. Full guide with scripts and a calculator below

rvaniaaa @rvaniaaaa

a week ago

x.com/i/article/2109…

4 14 97 149K 230

@cyrilXBT The routing split is the easy part. The real win is the critic loop — a model that tries to break your output before it ships. That's basically a watcher for code. Do you gate the merge on the critic failing to reproduce, or just flag it?

@mirku21 the /workflows point is the real one. sweeps don't need a model on every step — once the shape is known, freeze it to code and let the cheap tier only handle the judgment call. deterministic where you can, model where you must.

Opus 5.5 being “2x more expensive” is one of those facts that dies the second you run a long agent loop [the loop changes the math] 1. at 20K context, Opus is still 1.88x Sonnet per turn 2. at 150K, that falls to 1.49x. by 400K, it is only 1.26x 3. the reason is cache. both models pay the same $0.20 / 1M for cache reads, so old context stops widening the bill 4. the ugly part is switching late. drag 300K of failed context into Opus and you spend $1.50 just rebuilding cache 5. catch the failure around turn 6 and hand off 20K instead. that costs about $0.10 at 150K, 30 Sonnet turns cost $1.76 20 Opus turns cost $1.75 the real optimization is not “use the cheaper model” it is knowing when to stop paying for the wrong one ↓

Gipp 🦅 @gippp69

2 weeks ago

x.com/i/article/2106…

29 14 238 290K 263